Machine Learning Models predicting Decompensation in Cirrhosis

Abstract

Background and Aims: Decompensation of cirrhosis significantly decreases survival, thus, prevention of complications is paramount. We used machine learning techniques to identify parameters predicting decompensation.

Methods: Several machine learning techniques were applied to the INCA trial database containing pro- and retrospective data from 983 patients. Laboratory, clinical, and genetic data were analysed. After performing hierarchical clustering, Permutation Feature Importance was used to evaluate the impact of parameters on the prediction of decompensation.

Results: Achieving an accuracy of 81.6% on training and 70.5% on test data, Random Forests were best for retrospective prediction. In prospective assessment, Support Vector Machines performed best with an accuracy of 78.6% and 73.8%, respectively. Permutation Feature Importance demonstrated that baseline albumin and bilirubin levels and maximum bilirubin were the highest ranked parameters associated with former decompensation. In the prospective analysis, the maximum bilirubin value and the baseline values of sodium and albumin were ranked highest. In addition to the parameters of established scores, NOD2 genotype and inflammatory markers were highly ranked.

Conclusions: Laboratory parameters, genetic variants and infections can help to predict the risk of cirrhosis decompensation. This proof-of-concept study adds data for the future development of advanced models to identify patients at risk.